Research Of Dynamic Facial Expression Recognition Based On Deep Learning

In the development of human society,social communication has always been a relatively important part,and facial expressions can express the emotional state information of the human heart,which is of great significance to social interpersonal relationships.Therefore,the research on facial expression recognition has a strong practical significance,dynamic facial expression recognition has wide application,such as human-computer interaction,emotional mining,fatigue driving detection and other fields.This paper makes a detailed investigation of the research related to dynamic facial expression recognition with reference to a large number of domestic and foreign literatures and finds that there are two main methods for summarizing facial expression recognition: traditional machine learning methods and methods based on deep learning.The traditional machine learning methods are divided into four steps: image preprocessing,face alignment,facial feature extraction and expression classification.The feature extraction of images in traditional machine learning methods is the most important.This paper analyzes the feature extraction algorithms such as classical ASM,LDP and SIFT transform,and the classic expression classification algorithm SVM.Although most of the methods based on machine learning are mature,when dealing with dynamic expression recognition problems,because that ignore the temporal characteristics of the image by considering only the spatial features of the extracted image,so there are still problems of low accuracy,poor generalization ability,and complex feature engineering.In contrast,deep learning based methods have less these problems and the recognition effect is good.In order to solve the problem of temporal feature and spatial feature extraction and fusion in the field of dynamic facial expression recognition,this paper proposed a three-dimensional deep convolution residual dynamic temporal neural network to identify dynamic facial expressions.The neural network is composed of a Stem layer,a 3D Inception-ResNets structure,a GRU layer,a Dropout layer,an Island layer and a Softmax layer and can capture spatial relationships in facial expression images and temporal relationships between different face frames.In the dynamic facial expression recognition model proposed in this paper,traditional image preprocessing techniques such as gradation transformation and geometric transformation and face alignment algorithm are adopted,and the effect is good.In order to increase the contribution of the important face component in the expression recognition,in addition to inputting the facial expression frame data,important feature point information of the facial expression was extracted and input into the neural network.A novel Island loss function was introduced.Specifically,the island loss was designed to reduce the intra-class variations,while enlarging the inter-class differences simultaneously.This could better ignore the effects of facial expression variability and sensitivity,and achieve higher recognition accuracy?time efficiency and generalization ability.The paper carried out experiments on three public data sets CK+,AFEW and MMI in subject-independent and cross-database tasks.The proposed network is better than the current mainstream dynamic expression recognition method in accuracy.